src/pipelines.js (40 lines): - line 296: // TODO: Use softmax tensor function - line 430: // TODO add option to keep special tokens? - line 445: // TODO: Add support for start and end - line 516: // TODO: add support for `return_special_tokens_mask` - line 581: // TODO add start and end? - line 765: // TODO update generation config - line 775: // TODO: move to Translation pipeline? - line 1318: // TODO: Provide warning to the user that they might be using model which was not exported - line 1763: // TODO use kwargs - line 1836: // TODO support different stride_length_s (for left and right) - line 1882: // TODO: Right now we only get top beam - line 2244: target_sizes ?? imageSizes, // TODO FIX? - line 2668: // TODO: support separate threshold values - line 2845: // TODO: Find a better way for `pipeline` to set the default vocoder - line 3062: // TODO: replace with original - line 3073: // TODO: replace with original - line 3084: // TODO: replace with original - line 3096: // TODO: replace with original - line 3107: // TODO: replace with original - line 3118: // TODO: replace with original - line 3129: // TODO: replace with original - line 3140: // TODO: replace with original - line 3151: // TODO: replace with original - line 3162: // TODO: replace with original - line 3174: // TODO: replace with original - line 3186: // TODO: replace with original - line 3198: // TODO: replace with original - line 3210: // TODO: replace with original - line 3223: // TODO: replace with original - line 3236: // TODO: replace with original - line 3259: // TODO: replace with original - line 3272: // TODO: replace with original - line 3284: // TODO: replace with original - line 3296: // TODO: replace with original - line 3308: // TODO: replace with original - line 3320: // TODO: replace with original - line 3333: // TODO: replace with original - line 3344: // TODO: replace with original - line 3353: // TODO: Add types for TASK_ALIASES - line 3357: // "vqa": "visual-question-answering", // TODO: Add src/models.js (26 lines): - line 1334: // TODO: add min_tokens_to_keep - line 1338: // TODO: add min_tokens_to_keep - line 1604: // TODO: update decoder attention mask if the model requires it - line 1703: // Custom logic (TODO: move to Musicgen class) - line 1775: // TODO: Update to align with HF transformers' implementation - line 1833: // TODO: - line 1842: // TODO make > numInputs - line 1852: // TODO: when we do, we simply copy past key values and accumulate into single large tensor - line 1897: // TODO: If branching, use previous beam as a starting point - line 1903: // TODO: Support beam search - line 1928: // TODO: ensure all_input_ids is padded correctly... - line 1936: // TODO: - line 2254: // TODO: Add RoFormerForCausalLM and RoFormerForMultipleChoice - line 2340: // TODO add ElectraForPreTraining - line 3362: // TODO: Implement language detection - line 3507: // TODO: Improve batch processing - line 3846: // TODO: needed? - line 3993: // TODO: needed? - line 5181: // TODO: MobileViTForSemanticSegmentation - line 5196: // TODO: MobileViTV2ForSemanticSegmentation - line 6576: // output_cross_attentions = false, // TODO add - line 6641: // cross_attentions: null, // TODO add - line 6977: // TODO: assert `inferred` is an integer - line 7168: // TODO support re-using PKVs for input_ids.dims[1] !== 1 - line 7978: // TODO: Do not add new models here - line 8125: // TODO: Refactor to allow class to specify model src/tokenizers.js (11 lines): - line 124: let regex = pattern.Regex.replace(/\\([#&~])/g, '$1'); // TODO: add more characters to this list if necessary - line 1540: // TODO support all behaviours (config.behavior) - line 1712: // TODO use all of config: add_prefix_space, trim_offsets - line 2133: // TODO move to base class (like HF) - line 2161: // TODO add spaces_between_special_tokens and clean_up_tokenization_spaces options - line 2391: // TODO: detect when a different `this.charsmap` is used. - line 2620: // TODO: save this to the decoder when exporting? - line 3505: // TODO: Do the same for pair? - line 3605: // TODO add support for `return_language` - line 3712: // TODO Handle when language is different from the previous - line 4065: // TODO: implement offsets src/utils/tensor.js (9 lines): - line 56: // FIXME: ONNXTensor declares dims as readonly so one needs to use the constructor() if dims change. - line 113: // TODO allow setting of data - line 467: // TODO: implement transpose. For now (backwards compatibility), it's just an alias for permute() - line 625: // TODO validate inputs - line 654: // TODO: validate dims - line 941: // TODO: support scale_factor - line 1209: // TODO do validation of shapes - line 1271: // TODO do validation of shapes - line 1437: // TODO: support other dtypes src/utils/audio.js (3 lines): - line 301: // TODO warn if there is a zero row - line 625: // TODO: What if `mel_filters` is null? - line 634: // TODO: Make `mel_filters` a Tensor during initialization src/utils/hub.js (3 lines): - line 277: // TODO in future, cache the response? - line 363: // TODO add the rest? - line 541: // TODO in future, cache the response? src/base/image_processors_utils.js (3 lines): - line 377: // TODO add `should_fuse` - line 769: // TODO: Implement this - line 958: // TODO: src/models/sam/image_processing_sam.js (3 lines): - line 37: // TODO: add support for 2D input_points - line 125: // TODO allow user to use preprocessed images - line 238: // TODO: Implement src/backends/onnx.js (3 lines): - line 92: // TODO: Only push supported providers (depending on available hardware) - line 188: // TODO: Add support for loading WASM files from cached buffer when we upgrade to onnxruntime-web@1.19.0 - line 205: // TODO: Update this when allowing non-WASM backends. scripts/convert.py (2 lines): - line 300: # TODO allow user to specify vocoder path - line 384: # TODO: Enable once https://github.com/huggingface/optimum/pull/1552 is merged src/models/wav2vec2/feature_extraction_wav2vec2.js (2 lines): - line 11: // TODO support batch? - line 37: // TODO: allow user to pass in attention mask src/models/vitpose/image_processing_vitpose.js (2 lines): - line 23: // TODO: - line 60: // TODO: Determine best offsets src/models/detr/image_processing_detr.js (2 lines): - line 27: // TODO support differently-sized images, for now assume all images are the same size. - line 28: // TODO support different mask sizes (not just 64x64) scripts/float16.py (1 line): - line 184: # TODO: Ideally, "Cast" nodes should not be here, for the following reasons: src/utils/image.js (1 line): - line 389: // TODO use `resample` in browser environment src/base/processing_utils.js (1 line): - line 154: // TODO: src/utils/dtypes.js (1 line): - line 7: // TODO: Use the adapter from `env.backends.onnx.webgpu.adapter` to check for `shader-f16` support, src/generation/logits_process.js (1 line): - line 111: // // TODO: convert to `new Map(forced_decoder_ids)` src/generation/stopping_criteria.js (1 line): - line 100: // TODO: add MaxTimeCriteria src/utils/maths.js (1 line): - line 21: // TODO use mode and align_corners src/ops/registry.js (1 line): - line 43: // TODO: Allow for multiple execution providers scripts/quantize.py (1 line): - line 57: # TODO: "SimplifiedLayerNormalization", "SkipSimplifiedLayerNormalization" scripts/extra/marian.py (1 line): - line 33: 'precompiled_charsmap': None # TODO add this scripts/extra/clap.py (1 line): - line 1: # TODO: Enable once https://github.com/huggingface/optimum/pull/1552 is merged src/models/phi3_v/processing_phi3_v.js (1 line): - line 36: // TODO: Check if this affects performance, since it looks like a bug in the original implementation src/models/qwen2_vl/processing_qwen2_vl.js (1 line): - line 50: // TODO: ...videos_inputs, src/models/clap/feature_extraction_clap.js (1 line): - line 71: // TODO implement fusion strategy src/models/auto/processing_auto.js (1 line): - line 46: // TODO: first check for processor.json src/utils/video.js (1 line): - line 58: // TODO: Support efficiently loading all frames using the WebCodecs API. src/models/ultravox/processing_ultravox.js (1 line): - line 18: // TODO: Support batched inputs